> Eliezer S. Yudkowsky wrote:
> > Do you know that there's an STM connected to the Internet? I
> > don't know about protein-synthesis machines being connected to
> > the Internet - but if you're willing to go through snail-mail,
> > you can write a DNA sequence, email it to someone who mails you
> > the DNA, then send the DNA to a protein synthesizer in, what,
> > about a week? Probably less if you're an AI and you're willing
> > to pay for FedEx overnight.
> > Ever wonder if there's a back door into Zyvex's laboratory equipment
> > somewhere, with a password only an AI could crack? That's what
> > *I'd* do.
Doug Jones wrote:
> That's certainly interesting, but last I checked, nobody was even
> remotely close to building a microprocessor out of DNA and/or
> Provided one had some sort of marginally functional design for a
> RISC made of protein, it would of course be possible to put it
> through some heavy evolution to improve the design. This would
> however turn system architecture from an engineering discipline into
> a science- once evolving a processor that can run a benchmark
> program very well, the researchers would have to figure out what
> this beast *was*... and given algernon's law, the evolved protein
> processor would do the benchmark well, but anything else
> In the absence of a quick-turn hardware capability, my point stands:
> a singularity driven by an expanding AI seed would be bottlenecked
> by hardware.
What if you gave the AI online access to a company which built custom designed hardware, and an unlimited bank account (not all that unlikely - whoever builds it likely has some fair access to resources). It examines its own hardware (description kindly given to it by benefactor), redesigns it and sends the new design to custom hardware company X, and pays (money donated by same benefactor), new hardware is delivered, installed, AI moved to new hardware (thanks again benefactor), and the AI continues the process using the last design as the next step forward (it is now smarter than it was due to software and hardware enhancements created by its previous dumber self).
Slower than software self-modification, granted, but possible.
If you've got unlimited money, give the AI direct control over a factory (fully automated? Maybe not, could employ humans) which can implement it's new hardware designs, then set up the hardware and copy the AI across. If the construction process is too slow, the AI can turn it's attention to redesigning the factory to improve things, get the redesigned factory built, and continue.
Certainly you could give an intelligent AI it's own corporation, with a starting bank balance; it could trade to make money, and use the money to enhance itself + continue trade, raking in more money as it went because it would be getting smarter and so getting better at capitalism. It could employ people to do human type stuff (tricky physical things that monkeys are better at than robots), but essentially would be independent, and perhaps could design better robots if it really needed to break the link.